Scalable, Fine Grain, Parallelization of the Car-Parrinello ab initio Molecular Dynamics Method

PPL Technical Report 2005

Publication Type: Paper

Repository URL:

Download:
[BIB]

Abstract

Many important problems in material science, chemistry, solid-state
physics, and biophysics necessitate a first principles or ab initio
based molecular modeling approach. That is, atomic forces generated
from an energy function that explicitly includes electrons are
required to explore non-trivial, technologically interesting
systems. A first principles technique that has proved to be
particularly efficient and useful is the {\em Car-Parrinello ab
initio molecular dynamics method} (CPAIMD). This computationally
intensive method which is typically applied to study systems
containing 100-1000s of atoms, has resisted attempts to achieve
parallel scaling beyond processor numbers equal to the number of
electronic states (100-1000 processors in system sizes of
interest). Indeed, CPAIMD computations involve a large number of
phases with complex dependencies, that lead to difficult
communication optimization and load balancing problems. These
phases include multiple concurrent sparse 3D-Fast-Fourier
Transform(3D-FFT) computations, non-square matrix multiplies and
few concurrent dense 3D-FFT computations. Using Charm++ and the
concept of virtualization, the CPAIMD phases are discretized into a
large number of virtual processors which are, in turn, mapped
flexibly onto physical processors by the Charm++ runtime system and
dynamically adjusted to achieve high performance. A benchmark with
32 water molecules (128 states) scaling to more than 1000
processors is given, setting a precedent for this problem.

TextRef

Sameer Kumar and Yan Shi and Eric Bohm and L. V. Kale, "Scalable, fine grain,
parallelization of the Car-Parrinello ab initio molecular dynamics method",
UIUC, Dept. of Computer Science, 2005.

People

Research Areas